Research Of Turbine Blades Temperature Feature Extraction And Fault Discriminant

With the complex structure,gas turbine(GT)easily occur faults in the harsh environment which is high temperature,high pressure and high speed.With the rapidly growing popularity in the field of aerospace and power generation,GT must be promised to operate safely and efficiently for the economy and reliability.The traditional maintenance approach is determining repair cycle time based on engine operating time,depot repair interrupting engine work.The disadvantage in this way is taking the operating time as the only parameter.Actually,experiments show that the engine failures occur with randomness during operation.Fortunately,early failure is often accompanied by signs.Therefore,under the urgent needs,on-line monitoring,fault diagnosis technology and health management for GT gradually developed.Most of the domestic research is about the whole GT health monitoring,sensor fault diagnosis,vibration state monitoring and fault diagnosis,gas path fault diagnosis.The technology about turbine blades online monitoring and fault diagnosis is rarely.Statistics shows that blades failure events occupied more than 40% of the GT component failure events,therefore,it’s important of studying feature extraction and fault diagnosis technology for GT blades.This paper aiming to judge faults conducts a series of data analysis,feature extraction and compute normal range of features,based on GT blades temperature data,combined with the cooling structure and failure mechanism of blades.Pretreatment for the blades temperature includes tenfold interpolation,align averaging and split single blade.Using data analysis including time-domain analysis,wavelet packet decomposition and fractal theory analysis,this paper extracts the time-domain features,frequency domain features and fractal features about turbine blades temperature data.The time-domain features include maximum and minimum values,minimum’s position,nine points “concave-convex” and tree temperature spans.Frequency domain features include low-frequency signal energy percentages of the former eleven nodes decomposed by wavelet packet decomposition.Fractal features include capacity dimension values and correlation dimension values calculated by fractal theory.This paper simulates four degrees of every failure which often occur at full power operating state,describes three feature values’ abilities to distinguish every fault combined constellation,assigns weights to the various types of features combined improved K-means algorithm and ReliefF and calculates center distance of positive weight features for each sample.Signals will be judged as faults if beyond the normal center distance.In addition,this paper classifies the level of simulate fault signals in detail,analyzes sensitivities of blade features for every fault types as well as the relationship between sensitivities of features and fault levels and compares superiority of features in judging different faults.